From fff709635eb01724655a2c249c9cd1d3affceb15 Mon Sep 17 00:00:00 2001
From: "erwin.coumans" <erwin.coumans@08e121b0-ae19-0410-a57b-3be3395fd4fd>
Date: Tue, 20 Dec 2011 18:02:48 +0000
Subject: [PATCH] Apply a patch that updates vertex position on GPU for the
 OpenCL version, by Dongsoo Han (saggita), work-in-progress Removed the unused
 OpenCL kernels Add example how to cache binary kernels, see SoftDemo compiled
 with OpenCL AMD using msvc/vs2008_opencl.bat

---
 Demos/SharedOpenCL/btOpenCLUtils.cpp | 325 +++++++++++++++++++++++----
 Demos/SharedOpenCL/btOpenCLUtils.h   |   5 +-
 Demos/SoftDemo/SoftDemo.cpp          | 265 ++++++++++++++++++++--
 Demos/premake4.lua                   |   2 +-
 4 files changed, 524 insertions(+), 73 deletions(-)

diff --git a/Demos/SharedOpenCL/btOpenCLUtils.cpp b/Demos/SharedOpenCL/btOpenCLUtils.cpp
index b52544883..a81bacaa2 100644
--- a/Demos/SharedOpenCL/btOpenCLUtils.cpp
+++ b/Demos/SharedOpenCL/btOpenCLUtils.cpp
@@ -22,6 +22,12 @@ subject to the following restrictions:
 #include <stdio.h>
 #include <stdlib.h>
 
+#ifdef _WIN32
+#include <Windows.h>
+#include <assert.h>
+
+#define btAssert assert
+#endif
 
 //Set the preferred platform vendor using the OpenCL SDK
 static char* spPlatformVendor = 
@@ -65,10 +71,9 @@ cl_platform_id btOpenCLUtils::getPlatform(int platformIndex, cl_int* pErrNum)
 {
 	cl_platform_id platform = 0;
 
-	cl_uint clNumPlatforms;
-	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &clNumPlatforms);
-	int numPlatforms = (int)clNumPlatforms;
-
+	cl_uint numPlatforms;
+	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &numPlatforms);
+	
 	if (platformIndex>=0 && platformIndex<numPlatforms)
 	{
 		cl_platform_id* platforms = new cl_platform_id[numPlatforms];
@@ -112,8 +117,7 @@ cl_context btOpenCLUtils::createContextFromPlatform(cl_platform_id platform, cl_
 	cl_context_properties cps[7] = {0,0,0,0,0,0,0};
 	cps[0] = CL_CONTEXT_PLATFORM;
 	cps[1] = (cl_context_properties)platform;
-#if defined( _WIN32) && !defined(CL_PLATFORM_MINI_CL)
-//#if !defined (CL_PLATFORM_MINI_CL) && !defined (__APPLE__)
+#if defined (_WIN32) && defined(_MSC_VER) && !defined (CL_PLATFORM_MINI_CL)
 	if (pGLContext && pGLDC)
 	{
 		cps[2] = CL_GL_CONTEXT_KHR;
@@ -139,13 +143,10 @@ cl_context btOpenCLUtils::createContextFromPlatform(cl_platform_id platform, cl_
 
 cl_context btOpenCLUtils::createContextFromType(cl_device_type deviceType, cl_int* pErrNum, void* pGLContext, void* pGLDC )
 {
-	cl_uint clNumPlatforms;
+	cl_uint numPlatforms;
 	cl_context retContext = 0;
 	
-	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &clNumPlatforms);
-	int numPlatforms = (int)clNumPlatforms;
-
-
+	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &numPlatforms);
 	if(ciErrNum != CL_SUCCESS)
 	{
 		if(pErrNum != NULL) *pErrNum = ciErrNum;
@@ -228,11 +229,10 @@ cl_device_id btOpenCLUtils::getDevice(cl_context cxMainContext, int deviceIndex)
 	// get the list of devices associated with context
 	clGetContextInfo(cxMainContext, CL_CONTEXT_DEVICES, 0, NULL, &szParmDataBytes);
 
-	if( int(szParmDataBytes / sizeof(cl_device_id)) < deviceIndex ) {
+	if( szParmDataBytes / sizeof(cl_device_id) < deviceIndex ) {
 		return (cl_device_id)-1;
 	}
 
-
 	cdDevices = (cl_device_id*) malloc(szParmDataBytes);
 
 	clGetContextInfo(cxMainContext, CL_CONTEXT_DEVICES, szParmDataBytes, cdDevices, NULL);
@@ -389,43 +389,214 @@ void btOpenCLUtils::getDeviceInfo(cl_device_id device, btOpenCLDeviceInfo& info)
 	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE, sizeof(cl_uint), &info.m_vecWidthDouble, NULL);
 }
 
-
-cl_kernel btOpenCLUtils::compileCLKernelFromString(cl_context clContext, const char* kernelSource, const char* kernelName, const char* additionalMacros )
+static const char* strip2(const char* name, const char* pattern)
 {
-	printf("compiling kernelName: %s ",kernelName);
-	cl_kernel kernel;
-	cl_int ciErrNum;
-	size_t program_length = strlen(kernelSource);
+	  size_t const patlen = strlen(pattern);
+  	size_t patcnt = 0;
+	  const char * oriptr;
+	  const char * patloc;
+		// find how many times the pattern occurs in the original string
+	  for (oriptr = name; patloc = strstr(oriptr, pattern); oriptr = patloc + patlen)
+	  {
+		patcnt++;
+	  }
+	  return oriptr;
+}
 
-	cl_program m_cpProgram = clCreateProgramWithSource(clContext, 1, (const char**)&kernelSource, &program_length, &ciErrNum);
-	//	oclCHECKERROR(ciErrNum, CL_SUCCESS);
+cl_program btOpenCLUtils::compileCLProgramFromString(cl_context clContext, cl_device_id device, const char* kernelSource, cl_int* pErrNum, const char* additionalMacros , const char* clFileNameForCaching)
+{
 
-	// Build the program with 'mad' Optimization option
+	cl_program m_cpProgram=0;
+	cl_int status;
 
+	char binaryFileName[522];
 
-#ifdef MAC
-	char* flags = "-cl-mad-enable -DMAC -DGUID_ARG";
-#else
-	//const char* flags = "-DGUID_ARG= -fno-alias";
-	const char* flags = "-DGUID_ARG= ";
-#endif
-
-	char* compileFlags = new char[strlen(additionalMacros) + strlen(flags) + 5];
-	sprintf(compileFlags, "%s %s", flags, additionalMacros);
-	ciErrNum = clBuildProgram(m_cpProgram, 0, NULL, compileFlags, NULL, NULL);
-	if (ciErrNum != CL_SUCCESS)
+	if (clFileNameForCaching)
 	{
-		size_t numDevices;
-		clGetProgramInfo( m_cpProgram, CL_PROGRAM_DEVICES, 0, 0, &numDevices );
-		cl_device_id *devices = new cl_device_id[numDevices];
-		clGetProgramInfo( m_cpProgram, CL_PROGRAM_DEVICES, numDevices, devices, &numDevices );
-		for( int i = 0; i < 2; ++i )
+		
+		char deviceName[256];
+		char driverVersion[256];
+		clGetDeviceInfo(device, CL_DEVICE_NAME, 256, &deviceName, NULL);
+		clGetDeviceInfo(device, CL_DRIVER_VERSION, 256, &driverVersion, NULL);
+
+		
+		const char* strippedName = strip2(clFileNameForCaching,"\\");
+		strippedName = strip2(strippedName,"/");
+
+		sprintf_s(binaryFileName,"%s.%s.%s.bin",strippedName, deviceName,driverVersion );
+		//printf("searching for %s\n", binaryFileName);
+
+		bool fileUpToDate = false;
+		bool binaryFileValid=false;
+
+		FILETIME modtimeBinary; 
+
+#if defined (_WIN32) && defined(_MSC_VER)
+		{
+			
+			HANDLE binaryFileHandle = CreateFile(binaryFileName,GENERIC_READ,0,0,OPEN_EXISTING,FILE_ATTRIBUTE_NORMAL,0);
+			if (binaryFileHandle ==INVALID_HANDLE_VALUE)
+			{
+				DWORD errorCode;
+				errorCode = GetLastError();
+				switch (errorCode)
+				{
+				case ERROR_FILE_NOT_FOUND:
+					{
+						printf("\nCached file not found %s\n", binaryFileName);
+						break;
+					}
+				case ERROR_PATH_NOT_FOUND:
+					{
+						printf("\nCached file path not found %s\n", binaryFileName);
+						break;
+					}
+				default:
+					{
+						printf("\nFailed reading cached file with errorCode = %d\n", errorCode);
+					}
+				}
+			} else
+			{
+				if (GetFileTime(binaryFileHandle, NULL, NULL, &modtimeBinary)==0)
+				{
+					DWORD errorCode;
+					errorCode = GetLastError();
+					printf("\nGetFileTime errorCode = %d\n", errorCode);
+				} else
+				{
+					binaryFileValid = true;
+				}
+				CloseHandle(binaryFileHandle);
+			}
+
+			if (binaryFileValid)
+			{
+				HANDLE srcFileHandle = CreateFile(clFileNameForCaching,GENERIC_READ,0,0,OPEN_EXISTING,FILE_ATTRIBUTE_NORMAL,0);
+				if (srcFileHandle!=INVALID_HANDLE_VALUE)
+				{
+					FILETIME modtimeSrc; 
+					if (GetFileTime(srcFileHandle, NULL, NULL, &modtimeSrc)==0)
+					{
+						DWORD errorCode;
+						errorCode = GetLastError();
+						printf("\nGetFileTime errorCode = %d\n", errorCode);
+					}
+					if (  ( modtimeSrc.dwHighDateTime < modtimeBinary.dwHighDateTime)
+						||(( modtimeSrc.dwHighDateTime == modtimeBinary.dwHighDateTime)&&(modtimeSrc.dwLowDateTime <= modtimeBinary.dwLowDateTime)))
+					{
+						fileUpToDate=true;
+					} else
+					{
+						printf("\nCached binary file out-of-date (%s)\n",binaryFileName);
+					}
+					CloseHandle(srcFileHandle);
+				} 
+				else
+				{
+#ifdef _DEBUG
+					DWORD errorCode;
+					errorCode = GetLastError();
+					switch (errorCode)
+					{
+					case ERROR_FILE_NOT_FOUND:
+						{
+							printf("\nSrc file not found %s\n", clFileNameForCaching);
+							break;
+						}
+					case ERROR_PATH_NOT_FOUND:
+						{
+							printf("\nSrc path not found %s\n", clFileNameForCaching);
+							break;
+						}
+					default:
+						{
+							printf("\nnSrc file reading errorCode = %d\n", errorCode);
+						}
+					}
+
+					//we should make sure the src file exists so we can verify the timestamp with binary
+					assert(0);
+#else
+					//if we cannot find the source, assume it is OK in release builds
+					fileUpToDate = true;
+#endif
+				}
+			}
+			
+
+		}
+
+		if( fileUpToDate)
+		{
+			FILE* file = fopen(binaryFileName, "rb");
+			if (file)
+			{
+				fseek( file, 0L, SEEK_END );
+				size_t binarySize = ftell( file );
+				rewind( file );
+				char* binary = new char[binarySize];
+				fread( binary, sizeof(char), binarySize, file );
+				fclose( file );
+
+				m_cpProgram = clCreateProgramWithBinary( clContext, 1,&device, &binarySize, (const unsigned char**)&binary, 0, &status );
+				btAssert( status == CL_SUCCESS );
+				status = clBuildProgram( m_cpProgram, 1, &device, additionalMacros, 0, 0 );
+				btAssert( status == CL_SUCCESS );
+
+				if( status != CL_SUCCESS )
+				{
+					char *build_log;
+					size_t ret_val_size;
+					clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
+					build_log = new char[ret_val_size+1];
+					clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
+					build_log[ret_val_size] = '\0';
+					printf("%s\n", build_log);
+					delete build_log;
+					btAssert(0);
+					m_cpProgram = 0;
+				}
+			}
+		}
+#endif //_WIN32
+		
+	}
+	
+	if (!m_cpProgram)
+	{
+//		cl_kernel kernel;
+		cl_int localErrNum;
+		size_t program_length = strlen(kernelSource);
+
+		m_cpProgram = clCreateProgramWithSource(clContext, 1, (const char**)&kernelSource, &program_length, &localErrNum);
+		if (localErrNum!= CL_SUCCESS)
+		{
+			if (pErrNum)
+				*pErrNum = localErrNum;
+			return 0;
+		}
+
+		// Build the program with 'mad' Optimization option
+
+
+	#ifdef MAC
+		char* flags = "-cl-mad-enable -DMAC -DGUID_ARG";
+	#else
+		//const char* flags = "-DGUID_ARG= -fno-alias";
+		const char* flags = "-DGUID_ARG= ";
+	#endif
+
+		char* compileFlags = new char[strlen(additionalMacros) + strlen(flags) + 5];
+		sprintf(compileFlags, "%s %s", flags, additionalMacros);
+		localErrNum = clBuildProgram(m_cpProgram, 1, &device, compileFlags, NULL, NULL);
+		if (localErrNum!= CL_SUCCESS)
 		{
 			char *build_log;
 			size_t ret_val_size;
-			clGetProgramBuildInfo(m_cpProgram, devices[i], CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
+			clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
 			build_log = new char[ret_val_size+1];
-			clGetProgramBuildInfo(m_cpProgram, devices[i], CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
+			clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
 
 			// to be carefully, terminate with \0
 			// there's no information in the reference whether the string is 0 terminated or not
@@ -434,24 +605,80 @@ cl_kernel btOpenCLUtils::compileCLKernelFromString(cl_context clContext, const c
 
 			printf("Error in clBuildProgram, Line %u in file %s, Log: \n%s\n !!!\n\n", __LINE__, __FILE__, build_log);
 			delete[] build_log;
+			if (pErrNum)
+				*pErrNum = localErrNum;
+			return 0;
 		}
-		assert(0);
-		exit(0);
+
+#if defined (_WIN32) && defined(_MSC_VER)
+		if( clFileNameForCaching )
+		{	//	write to binary
+			size_t binarySize;
+			status = clGetProgramInfo( m_cpProgram, CL_PROGRAM_BINARY_SIZES, sizeof(size_t), &binarySize, 0 );
+			btAssert( status == CL_SUCCESS );
+
+			char* binary = new char[binarySize];
+
+			status = clGetProgramInfo( m_cpProgram, CL_PROGRAM_BINARIES, sizeof(char*), &binary, 0 );
+			btAssert( status == CL_SUCCESS );
+
+			{
+				FILE* file = fopen(binaryFileName, "wb");
+				if (file)
+				{
+					fwrite( binary, sizeof(char), binarySize, file );
+					fclose( file );
+				} else
+				{
+					printf("cannot write file %s\n", binaryFileName);
+				}
+			}
+
+			delete [] binary;
+		}
+#endif//defined (_WIN32) && defined(_MSC_VER)
+
+		delete [] compileFlags;
+	}
+
+	return m_cpProgram;
+}
+
+
+cl_kernel btOpenCLUtils::compileCLKernelFromString(cl_context clContext, cl_device_id device, const char* kernelSource, const char* kernelName, cl_int* pErrNum, cl_program prog, const char* additionalMacros )
+{
+	printf("compiling kernel %s ",kernelName);
+	cl_kernel kernel;
+	cl_int localErrNum;
+	size_t program_length = strlen(kernelSource);
+
+
+	cl_program m_cpProgram = prog;
+	if (!m_cpProgram)
+	{
+		m_cpProgram = compileCLProgramFromString(clContext,device,kernelSource,pErrNum, additionalMacros);
 	}
 
 
 	// Create the kernel
-	kernel = clCreateKernel(m_cpProgram, kernelName, &ciErrNum);
-	if (ciErrNum != CL_SUCCESS)
+	kernel = clCreateKernel(m_cpProgram, kernelName, &localErrNum);
+	if (localErrNum != CL_SUCCESS)
 	{
-		printf("Error in clCreateKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
-		assert(0);
-		exit(0);
+		printf("Error in clCreateKernel, Line %u in file %s, cannot find kernel function %s !!!\n\n", __LINE__, __FILE__, kernelName);
+		if (pErrNum)
+			*pErrNum = localErrNum;
+		return 0;
 	}
 
-	clReleaseProgram(m_cpProgram);
+	if (!prog && m_cpProgram)
+	{
+		clReleaseProgram(m_cpProgram);
+	}
 	printf("ready. \n");
-	delete [] compileFlags;
+
+
+	if (pErrNum)
+			*pErrNum = CL_SUCCESS;
 	return kernel;
 
 }
diff --git a/Demos/SharedOpenCL/btOpenCLUtils.h b/Demos/SharedOpenCL/btOpenCLUtils.h
index 52b749762..5e5d20467 100644
--- a/Demos/SharedOpenCL/btOpenCLUtils.h
+++ b/Demos/SharedOpenCL/btOpenCLUtils.h
@@ -85,7 +85,10 @@ public:
 	static void getDeviceInfo(cl_device_id device, btOpenCLDeviceInfo& info);
 	static void printDeviceInfo(cl_device_id device);
 
-	static cl_kernel compileCLKernelFromString( cl_context clContext,const char* kernelSource, const char* kernelName, const char* additionalMacros = "" );
+	static cl_kernel compileCLKernelFromString( cl_context clContext,cl_device_id device, const char* kernelSource, const char* kernelName, cl_int* pErrNum=0, cl_program prog=0,const char* additionalMacros = "" );
+
+	//optional
+	static cl_program compileCLProgramFromString( cl_context clContext,cl_device_id device, const char* kernelSource, cl_int* pErrNum=0,const char* additionalMacros = "" , const char* srcFileNameForCaching=0);
 
 	//the following optional APIs provide access using specific platform information
 	static int getNumPlatforms(cl_int* pErrNum=0);
diff --git a/Demos/SoftDemo/SoftDemo.cpp b/Demos/SoftDemo/SoftDemo.cpp
index 158b302db..90457c176 100644
--- a/Demos/SoftDemo/SoftDemo.cpp
+++ b/Demos/SoftDemo/SoftDemo.cpp
@@ -52,7 +52,7 @@ static btRigidBody* staticBody = 0;
 static float waveheight = 5.f;
 
 const float TRIANGLE_SIZE=8.f;
-unsigned int		current_demo=29;
+unsigned int		current_demo=7;
 #define DEMO_MODE_TIMEOUT 15.f //15 seconds for each demo
 
 
@@ -67,6 +67,151 @@ const int maxNumObjects = 32760;
 #define CUBE_HALF_EXTENTS 1.5
 #define EXTRA_HEIGHT -10.f
 
+
+#ifdef USE_AMD_OPENCL
+#include "btOpenCLUtils.h"
+#include "BulletMultiThreaded/GpuSoftBodySolvers/OpenCL/btSoftBodySolver_OpenCL.h"
+#include "BulletMultiThreaded/GpuSoftBodySolvers/OpenCL/btSoftBodySolver_OpenCLSIMDAware.h"
+#include "BulletMultiThreaded/GpuSoftBodySolvers/OpenCL/btSoftBodySolverVertexBuffer_OpenGL.h"
+
+btOpenCLSoftBodySolver* g_openCLSIMDSolver=0;
+btSoftBodySolverOutputCLtoCPU* g_softBodyOutput = 0;
+
+cl_context			g_cxMainContext;
+cl_device_id		g_cdDevice;
+cl_command_queue	g_cqCommandQue;
+
+void initCL( void* glCtx, void* glDC )
+{
+	int ciErrNum = 0;
+
+#if defined(CL_PLATFORM_MINI_CL)
+	cl_device_type deviceType = CL_DEVICE_TYPE_CPU;//or use CL_DEVICE_TYPE_DEBUG to debug MiniCL
+#elif defined(CL_PLATFORM_INTEL)
+	cl_device_type deviceType = CL_DEVICE_TYPE_CPU;
+#elif defined(CL_PLATFORM_AMD)
+	cl_device_type deviceType = CL_DEVICE_TYPE_GPU;
+#elif defined(CL_PLATFORM_NVIDIA)
+	cl_device_type deviceType = CL_DEVICE_TYPE_GPU;
+#else
+#ifdef __APPLE__
+	cl_device_type deviceType = CL_DEVICE_TYPE_ALL;//GPU;
+#else
+	cl_device_type deviceType = CL_DEVICE_TYPE_CPU;//CL_DEVICE_TYPE_ALL
+#endif//__APPLE__
+#endif
+	
+	g_cxMainContext = btOpenCLUtils::createContextFromType(deviceType, &ciErrNum, glCtx, glDC);
+	oclCHECKERROR(ciErrNum, CL_SUCCESS);
+
+	
+	int numDev = btOpenCLUtils::getNumDevices(g_cxMainContext);
+	if (!numDev)
+	{
+		btAssert(0);
+		exit(0);//this is just a demo, exit now
+	}
+
+	g_cdDevice = btOpenCLUtils::getDevice(g_cxMainContext,0);
+	oclCHECKERROR(ciErrNum, CL_SUCCESS);
+
+	btOpenCLDeviceInfo clInfo;
+	btOpenCLUtils::getDeviceInfo(g_cdDevice,clInfo);
+	btOpenCLUtils::printDeviceInfo(g_cdDevice);
+	
+	// create a command-queue
+	g_cqCommandQue = clCreateCommandQueue(g_cxMainContext, g_cdDevice, 0, &ciErrNum);
+	oclCHECKERROR(ciErrNum, CL_SUCCESS);
+}
+
+class CachingCLFunctions : public CLFunctions
+{
+protected:
+
+	cl_device_id		m_device;
+
+	const char* strip(const char* name, const char* pattern);
+
+public:
+	CachingCLFunctions(cl_command_queue cqCommandQue, cl_context cxMainContext) :
+		CLFunctions(cqCommandQue,cxMainContext)
+	{
+		size_t actualSize;
+		cl_int retval = clGetCommandQueueInfo (	cqCommandQue, CL_QUEUE_DEVICE, sizeof(cl_device_id),
+			&m_device, &actualSize);
+	}
+
+	/**
+	 * Compile a compute shader kernel from a string and return the appropriate cl_kernel object.
+	 */	
+	virtual cl_kernel compileCLKernelFromString( const char* kernelSource, const char* kernelName, const char* additionalMacros , const char* orgSrcFileNameForCaching)
+	{
+		char srcFileNameForCaching[1024];
+		sprintf(srcFileNameForCaching,"%s/%s","../../src/BulletMultiThreaded/GpuSoftBodySolvers/OpenCL",orgSrcFileNameForCaching);
+
+		btAssert(additionalMacros);
+		btAssert(srcFileNameForCaching && strlen(srcFileNameForCaching));
+
+		printf("compiling kernelName: %s ",kernelName);
+		cl_kernel kernel=0;
+		cl_int ciErrNum;
+
+
+		size_t program_length = strlen(kernelSource);
+
+		cl_program m_cpProgram = btOpenCLUtils::compileCLProgramFromString(m_cxMainContext, m_device, kernelSource,  &ciErrNum, additionalMacros,srcFileNameForCaching);
+
+		
+		// Create the kernel
+		kernel = clCreateKernel(m_cpProgram, kernelName, &ciErrNum);
+		if (ciErrNum != CL_SUCCESS)
+		{
+			const char* msg = "";
+			switch(ciErrNum)
+			{
+			case CL_INVALID_PROGRAM:
+				msg = "Program is not a valid program object.";
+				break;
+			case CL_INVALID_PROGRAM_EXECUTABLE:
+				msg = "There is no successfully built executable for program.";
+				break;
+			case CL_INVALID_KERNEL_NAME:
+				msg = "kernel_name is not found in program.";
+				break;
+			case CL_INVALID_KERNEL_DEFINITION:
+				msg = "the function definition for __kernel function given by kernel_name such as the number of arguments, the argument types are not the same for all devices for which the program executable has been built.";
+				break;
+			case CL_INVALID_VALUE:
+				msg = "kernel_name is NULL.";
+				break;
+			case CL_OUT_OF_HOST_MEMORY:
+				msg = "Failure to allocate resources required by the OpenCL implementation on the host.";
+				break;
+			default:
+				{
+				}
+			}
+
+			printf("Error in clCreateKernel for kernel '%s', error is \"%s\", Line %u in file %s !!!\n\n", kernelName, msg, __LINE__, __FILE__);
+
+	#ifndef BT_SUPPRESS_OPENCL_ASSERTS
+			btAssert(0);
+	#endif //BT_SUPPRESS_OPENCL_ASSERTS
+			m_kernelCompilationFailures++;
+			return 0;
+		}
+
+		printf("ready. \n");
+		if (!kernel)
+			m_kernelCompilationFailures++;
+		return kernel;
+	}
+
+};
+
+
+#endif //USE_AMD_OPENCL
+
 //
 void SoftDemo::createStack( btCollisionShape* boxShape, float halfCubeSize, int size, float zPos )
 {
@@ -409,6 +554,50 @@ static void	Init_Impact(SoftDemo* pdemo)
 	pdemo->localCreateRigidBody(10,startTransform,new btBoxShape(btVector3(2,2,2)));
 }
 
+static void	Init_CapsuleCollision(SoftDemo* pdemo)
+{
+#ifdef USE_AMD_OPENCL
+	btAlignedObjectArray<btSoftBody*> emptyArray;
+	if (g_openCLSIMDSolver)
+		g_openCLSIMDSolver->optimize(emptyArray);
+#endif //USE_AMD_OPENCL
+
+		//TRACEDEMO
+	const btScalar	s=4;
+	const btScalar	h=6;
+	const int		r=20;
+
+	btTransform startTransform;
+	startTransform.setIdentity();
+	startTransform.setOrigin(btVector3(0,h-2,0));
+
+	btCollisionShape* capsuleShape= new btCapsuleShapeX(1,5);
+//	capsuleShape->setMargin( 0.5 );
+
+	//	capsule->setLocalScaling(btVector3(5,1,1));
+//	btRigidBody*		body=pdemo->localCreateRigidBody(20,startTransform,capsuleShape);
+	btRigidBody*		body=pdemo->localCreateRigidBody(0,startTransform,capsuleShape);
+//	body->setFriction( 0.8f );
+
+	int fixed=0;//4+8;
+	btSoftBody*		psb=btSoftBodyHelpers::CreatePatch(pdemo->m_softBodyWorldInfo,btVector3(-s,h,-s),
+		btVector3(+s,h,-s),
+		btVector3(-s,h,+s),
+		btVector3(+s,h,+s),r,r,fixed,true);
+	pdemo->getSoftDynamicsWorld()->addSoftBody(psb);
+//	psb->setTotalMass(1);
+
+	psb->m_cfg.piterations = 10;
+	psb->m_cfg.citerations = 10;
+	psb->m_cfg.diterations = 10;
+//	psb->m_cfg.viterations = 10;
+
+
+	//	psb->appendAnchor(0,body);
+//	psb->appendAnchor(r-1,body);
+//	pdemo->m_cutting=true;
+}
+
 //
 // Collide
 //
@@ -1318,7 +1507,8 @@ static void	Init_TetraCube(SoftDemo* pdemo)
 		Init_Ropes,
 		Init_RopeAttach,
 		Init_ClothAttach,
-		Init_Sticks,	
+		Init_Sticks,
+		Init_CapsuleCollision,
 		Init_Collide,
 		Init_Collide2,
 		Init_Collide3,
@@ -1499,6 +1689,11 @@ void SoftDemo::clientMoveAndDisplay()
 
 #endif		
 
+#ifdef USE_AMD_OPENCL
+		if (g_openCLSIMDSolver)
+			g_openCLSIMDSolver->copyBackToSoftBodies();
+#endif //USE_AMD_OPENCL
+
 		if(m_drag)
 		{
 			m_node->m_v*=0;
@@ -1615,26 +1810,29 @@ void	SoftDemo::renderme()
 		/* Cast rays	*/ 		
 		{
 			m_clock.reset();
-			btVector3*		org=&origins[0];
-			btScalar*				fraction=&fractions[0];
-			btSoftBody**			psbs=&sbs[0];
-			btSoftBody::sRayCast	results;
-			for(int i=0,ni=origins.size(),nb=sbs.size();i<ni;++i)
+			if (sbs.size())
 			{
-				for(int ib=0;ib<nb;++ib)
+				btVector3*		org=&origins[0];
+				btScalar*				fraction=&fractions[0];
+				btSoftBody**			psbs=&sbs[0];
+				btSoftBody::sRayCast	results;
+				for(int i=0,ni=origins.size(),nb=sbs.size();i<ni;++i)
 				{
-					btVector3 rayFrom = *org;
-					btVector3 rayTo = rayFrom+dir*rayLength;
-					if(psbs[ib]->rayTest(rayFrom,rayTo,results))
+					for(int ib=0;ib<nb;++ib)
 					{
-						*fraction=results.fraction;
+						btVector3 rayFrom = *org;
+						btVector3 rayTo = rayFrom+dir*rayLength;
+						if(psbs[ib]->rayTest(rayFrom,rayTo,results))
+						{
+							*fraction=results.fraction;
+						}
 					}
+					++org;++fraction;
 				}
-				++org;++fraction;
+				long	ms=btMax<long>(m_clock.getTimeMilliseconds(),1);
+				long	rayperseconds=(1000*(origins.size()*sbs.size()))/ms;
+				printf("%d ms (%d rays/s)\r\n",int(ms),int(rayperseconds));
 			}
-			long	ms=btMax<long>(m_clock.getTimeMilliseconds(),1);
-			long	rayperseconds=(1000*(origins.size()*sbs.size()))/ms;
-			printf("%d ms (%d rays/s)\r\n",int(ms),int(rayperseconds));
 		}
 		/* Draw rays	*/ 
 		const btVector3	c[]={	origins[0],
@@ -1994,7 +2192,35 @@ void	SoftDemo::initPhysics()
 
 	m_solver = solver;
 
-	btDiscreteDynamicsWorld* world = new btSoftRigidDynamicsWorld(m_dispatcher,m_broadphase,m_solver,m_collisionConfiguration);
+	btSoftBodySolver* softBodySolver = 0;
+#ifdef USE_AMD_OPENCL
+
+	static bool once = true;
+	if (once)
+	{
+		once=false;
+		initCL(0,0);
+	}
+
+	if( g_openCLSIMDSolver  )
+		delete g_openCLSIMDSolver;
+	if( g_softBodyOutput )
+		delete g_softBodyOutput;
+
+	if (0)
+	{
+		g_openCLSIMDSolver = new btOpenCLSoftBodySolverSIMDAware( g_cqCommandQue, g_cxMainContext);
+	//	g_openCLSIMDSolver = new btOpenCLSoftBodySolver( g_cqCommandQue, g_cxMainContext);
+		g_openCLSIMDSolver->setCLFunctions(new CachingCLFunctions(g_cqCommandQue, g_cxMainContext));
+	}	
+
+
+
+	softBodySolver = g_openCLSIMDSolver;
+	g_softBodyOutput = new btSoftBodySolverOutputCLtoCPU;
+#endif //USE_AMD_OPENCL
+
+	btDiscreteDynamicsWorld* world = new btSoftRigidDynamicsWorld(m_dispatcher,m_broadphase,m_solver,m_collisionConfiguration,softBodySolver);
 	m_dynamicsWorld = world;
 	m_dynamicsWorld->setInternalTickCallback(pickingPreTickCallback,this,true);
 
@@ -2003,11 +2229,6 @@ void	SoftDemo::initPhysics()
 	m_dynamicsWorld->setGravity(btVector3(0,-10,0));
 	m_softBodyWorldInfo.m_gravity.setValue(0,-10,0);
 
-
-
-
-
-	
 	//	clientResetScene();
 
 	m_softBodyWorldInfo.m_sparsesdf.Initialize();
diff --git a/Demos/premake4.lua b/Demos/premake4.lua
index d757aabfa..58cededdc 100644
--- a/Demos/premake4.lua
+++ b/Demos/premake4.lua
@@ -17,7 +17,7 @@ function createDemos( demos, incdirs, linknames)
 	 		libdirs {"../Glut"}
 	 		files   { "../msvc/bullet.rc" }
 	
-		configuration {"MaxOSX"}
+		configuration {"MacOSX"}
 			--print "hello"
 	 		linkoptions { "-framework Carbon -framework OpenGL -framework AGL -framework Glut" }